HBase: structured storage of sparse data for Hadoop

Transcription

1 HBase: structured storage of sparse data for Hadoop Jim Kellerman, Inc.

2 Topics Introduction, brief history Concepts Architecture Client API Tools Project Status Questions And Answers

3 Brief History 6/06 Mike Cafarella posts on Hadoop mailing lists 10-11/06 interest in Bigtable, recruit others, some design documents 01-02/07 resumes work Mike Cafarella provides initial code base 04/07 Michael Stack joins 10/07 First usable version of HBase in Hadoop release

4 Concepts Data addressed by row/column/version key Version can be a specified by client default is System.currentTimeMillis Data is stored column-oriented rather than row-oriented More space efficient nulls are free Better compression data is similar Updates lock entire row Don't need to update every column Locks never span rows

5 Concepts (contd.) Columns (aka column families): Values are byte[] Most options on per-column basis: # of versions, compression, bloom filters, maximum value length Fixed at table creation Can be added/dropped if table is off-line Family members can be created/deleted at any time

6 Concepts (contd.) Example: Web Crawl data

7 Concepts (contd.) Tables are stored in regions A region is a row range for the table [start-key:end-key) When regions get too big they are split The two new regions get ½ the row range of the parent region» One gets lower half of row range» One gets upper half of row range Splits are instantaneous Each column family is stored in an HStore Each region has one HStore per column family

8 Concepts (contd.) When a write occurs: It is written to a write ahead log It is cached in memory Periodically, the cache is written to disk, creating a new file in each HStore for the columns being flushed A compaction occurs when the number of files in an HStore exceeds a threshold» Compactions are done in the background Periodically, the log file is closed and a new one created Old log files are garbage collected

9 Concepts (contd.) 1) write 2) redo log roll log old redo log 3) cache persisted data t0 cache flush persisted data t1 new redo log compaction Merged persisted data

10 Concepts (contd.) When a read occurs: 1)Check for data in cache 2)Look for data in persisted data, from newest to oldest. read 1) cache 2 cache miss? y persisted data t1 3 found? n persisted data t0

11 Concepts (contd.) The location of all user regions is stored in the META region Each META row maps one user region Each META region can map about 64K regions All the META regions are mapped by one ROOT region ROOT and META can map about 8 x 10 9 user regions With current region size of 64MB, about bytes of data can be stored

12 Concepts (contd.) Root Region meta region1 key meta region2 key Meta Region1 user region1 key user region2 key Meta Region user regionn key User Region1 user row1 user row User Region2 User RegionN

13 Architecture There are three major components: Master server Region server Client API

14 Architecture: Master Server Assigns regions to region servers The ROOT region is assigned first Scans ROOT region to find META regions, assigns them to region servers Scans META regions to find user regions, assigns them to region servers Reassigns regions for load balancing or if region server fails Tells client where ROOT region is located

15 Architecture: Region Server Server threads handle client requests Main thread is master heart beat loop Other threads: Process long running operations resulting from master heart beat response Check to see if regions need to be split Check to see if cache needs to be flushed Check to see if log needs to be rolled One thread per client request (leases)

16 Architecture: HRegion One HRegion per table fragment (row range) In memory cache of recent writes One HStore object per column Finds and reads data from HDFS May manage many files (one is created per cache flush) Performs compaction if too many files Performs split operation for a single column

17 Client API Create new HTable object to open a table Client locates ROOT region from master Client reads (and caches) ROOT region to locate META servers Client reads (and caches) META information for the table being opened Client requests are sent directly to region servers. Master is not involved Administrative functions to manipulate tables and columns

18 Client API (contd.) Read Specific row/column pair, specific row all columns Most recent version Specific version N versions All versions Scan multiple rows Write All row mutations are atomic Can write multiple columns in single update

19 Client API Examples // Open table HTable table = new HTable(conf, tablename); // Storing data long writeid = table.startupdate(row); table.put(writeid, columnname1, bytes); table.put(writeid, columnname2, bytes); table.delete(writeid, columnname3); table.commit(writeid); // Reading data byte[] data = table.get(row, columnname1);

20 Client API Examples (contd.) // Open scanner HScannerInterface scanner = table.obtainscanner(cols, new Text()); try { SortedMap<Text, byte[]> values = new TreeMap<Text, byte[]>(); HStoreKey currentkey = new HStoreKey(); while (scanner.next(currentkey, values)) { // row: currentkey.getrow(), version: currentkey.gettimestamp() for (Map.Entry<Text, byte[]> e: values.entryset()) { // columnname: e.getkey(), value: e.getvalue() } } } finally { scanner.close(); }

21 Tools HBase shell Web Interface

22 Tools: HBase Shell durruti$./bin/hbase shell Hbase Shell, version. Copyright (c) 2007 by udanax, licensed to Apache Software Foundation. Type 'help;' for usage. Hbase> create table 'test' ('test'); Hbase> insert into 'test' ('test:test') values ('some old value') where row="test_row"; Hbase> select * from 'test'; Row Column Cell test_row test:test some old value Hbase>

23 Tools: Web Interface

24 Tools: Web Interface (contd.)

25 Tools: Web Interface (contd.)

26 Project Status First usable release of HBase included in Hadoop However data loss possible without Hadoop append support. To do: Build community: users, contributors Documentation and ease-of-use features Performance analysis ZooKeeper integration More Monitoring

27 Project Status (contd.) Several key contributions to date. Map/Reduce connector HBase shell Relational operators (in progress) Restructure so applications can access multiple tables simultaneously. Interested? Get involved! Contributions welcome! Follow lists and Jira

28 Questions? And Answers! References: Bigtable: A Distributed Storage System for Structured Data The HBase Wiki at Apache.org The #hbase IRC chat room at freenode.net The Hadoop mailing lists